The electric machines diagnosis techniques most commonly used nowadays are based on the analysis of the current (due to its noninvasive characteristic) through the Fourier Transform (FT). Its main drawback is that it cannot be used in applications constantly working in transient regime, such as windmills or electric vehicles, among other fields of growing interest. Since the late XX century to date, some techniques have been developed for the diagnosis in transient regimes; these techniques are basically based on obtaining the time evolution of the current harmonic components caused by the faults, which is achieved using time-frequency (t-f) transforms. Up to now, standard non optimized transforms have been applied for the fault diagnosis in electric machines (e.g., short time FT, wavelet transform) which enable the detection of some faulty components in certain zones of the t-f plane. On the other hand, there exist transforms with adaptive characteristics, whose analysis is adjusted to the analyzed signal (e.g., Matching Pursuit), that have not been used in the diagnosis field. Nevertheless, these transforms do not enable to focus in obtaining the faulty components, and consume too much computational time (weeks).

In the present thesis, a new methodology of t-f analysis is developed, optimized for the fault diagnosis in electric machines, via current analysis. The proposed methodology is developed taking into account the special features of the signal analyzed and the diagnosis objectives; these enables monitoring multiple faulty components throughout wide areas of the t-f plane consuming small computational times, which enables high reliability diagnosis. The development of the methodology comprises the following stages: (i) The faulty components evolutions in the t-f plane are characterized; the thesis is focused in the bar breakage and eccentricity diagnosis in induction motors (IM). With the aim of estimating these evolutions for each captured current avoiding the use of speed sensors, an original method is developed to obtain the slip evolution from the captured current, even in the presence of faults. (iii) The thesis proposes the detection of faulty components through the current correlation with a family of t-f atoms (functions whose energy is concentrated around a point of the plane). (iv) The Frequency B-Splines (FBS) are selected as the optimal atoms family to be used, since they are capable of generating results with resolutions as high as those obtained with the Gabor functions (which reach the highest energy concentration possible), but consuming lower computational times. (v) The adaptive slope transform is defined, which, on the contrary of the already mentioned transforms, enables choosing the resolution at each point of the t-f plane, being able to adapt the analysis to each current in order to optimally obtain the faulty components evolutions. To that end, the concept of atom slope is defined as the quotient between the dispersion of its energy in frequency and in time. The atom slope used determines the quotient between the resolutions obtained in time and in frequency. Next, the slope criterion is defined: in order to obtain the optimal t-f resolution, the slope to be used in each point of the plane has to be equal to the slope of the component evolution to be detected in that point. (vi) Finally, the problem of quantifying using this type of transforms is solved, and techniques for improving the components visual identification are proposed.

The diagnosis methodology has been experimentally validated through the detection of bar breakages in IM directly fed from the main and eccentricities in IM fed either from the main or from an inverter.